planetary rover visual motion estimation improvement for...
TRANSCRIPT
552
Planetary Rover Visual Motion Estimation improvement for Autonomous,
Intelligent, and Robust Guidance, Navigation and Control
J. Nsasi Bakambu *, C. Langley *, G. Pushpanathan *, W. James MacLean *, R. Mukherji *, E.
Dupuis **
* MDA Corporation, Brampton, Ontario, Canada
e-mail: Joseph.bakambu, chris.langley, giri.pushpanathan, [email protected];
** Space Exploration, Canadian Space Agency, Saint-Hubert, Quebec, Canada
e-mail: [email protected]
Abstract
This paper presents the Mojave Desert field test
results of an improved planetary rover visual motion
estimation technique for the Autonomous, Intelligent,
and Robust Guidance, Navigation, and Control for
Planetary Rovers (AIR-GNC). The main improvements
include: optimal use of different features from stereo-pair
images as visual landmarks, and the use of VME-based
feedback to close the path tracking loop. As well, a
long-range and wide FOV active 3D sensor was used to
extract long-range fixed landmarks for enabling visual
motion estimation observability, and thus improving the
accuracy of the VME. The field test, conducted in
relevant Mars-like terrains, under dramatically changing
weather and lighting conditions, shows good localization
accuracy on average. Moreover, the MDA developed
Enhanced IMU-corrected odometry was reliable and had
good accuracy in all test locations including in loose
sand dunes. These results are based on data collected
during 7.3 km of traverses, under both fully autonomous
and tele-operated control.
1 Introduction
One of the continuing challenges for future
unmanned exploration rover missions on the moon and
Mars is the ability to accurately estimate the rover’s
position and orientation. In the absence of an equivalent
to the Global Positioning System on earth, designers of
future system must exploit a combination of vehicle
odometry, inertial sensing and visual information to
obtain this localization information. Future missions
such as ESA’s ExoMars and NASA’s Mars Mobile
Science Laboratory rover require the rover to
autonomously traverse hundreds of meters to one km
daily at speeds of up to 100 m/h [1]. Equally
challenging is the need to localize the position of the
rover to an accuracy of between one and four percent of
distance traveled. Accurate localization is arguably the
most fundamental competence required for long range
autonomous navigation. In this context, MDA Space
Missions and the Canadian Space Agency (CSA) have
embarked on a project to further their capability for
visual motion estimation (VME) of planetary rovers,
which is described in this paper.
VME algorithms have recently seen a considerable
amount of interest from the planetary exploration rover
community as a solution for accurate localization. On the
Mars Exploration Rovers (MERs), VME was not
considered part of the main system for localization, but
was shown to work relatively well. In [2], JPL reported
that “Visual Odometry software has enabled precision
drives over distances as long as 8 m on slopes greater
than 20 degrees, and has made it possible to safely
traverse the loose sandy plains of Meridiani”. The
algorithm works by tracking features (Harris corners) in
a stereo image pair from one frame to the next
(frame-to-frame). Thus, the problem is one of
determining “the change in position and attitude for two
pairs of stereo images by propagating uncertainty in a 3D
to 3D pose estimation formulation using maximum
likelihood estimation”. The evaluation tests conducted at
the JPL Marsyard and Johnson Valley, California showed
that the absolute position errors were less than 2.5% over
the 24 meter Marsyard course, and less than 1.5% over
the 29 meter Johnson Valley course. The rotation error
was less than 5.0 degrees in each case.
LAAS/CNRS (Laboratoire d'Architecture et
d'Analyse des Systèmes/Centre National de la Recherche
Scientifique) VME is based on the frame-to-frame pixel
tracking method. Landmarks are extracted from images
by finding points of interest identified by image intensity
gradients. The test results using the Lama rover [3]
showed an error of 4% on a 25 meter traverse. After
improvement of the algorithm in [4], the overall error of
2% on a 70 m traverse was maintained.
A survey of the literature [5][6][7] shows that the
i-SAIRAS 2010August 29-September 1, 2010, Sapporo, Japan
553
state-of-the-art in localization has yet to meet the
ExoMars localization accuracy requirement of 1% of the
traveled distance, without the extensive use of
post-processing methods (for example, bundle
adjustment).
Previous work at MDA has shown that a full
simultaneous localization and mapping (SLAM)
approach yields good accuracy, but has a high
computational burden, forcing lower commanded speeds
for the vehicle [8]. Conversely, if no map is maintained,
a frame-to-frame technique has been shown to operate
with acceptable speed but at the cost of decreased
accuracy. MDA proposed a multi-frame VME algorithm
that attempts to obtain a suitable balance between
accuracy and speed. The results in hardware indicate
that traverses of greater than 200 m are possible to an
accuracy between one and four percent in planetary
conditions. Our approach uses 3D odometry and a
stereo-pair to identify visual landmarks by using the
Scale Invariant Feature Transform (SIFT). While the
result is commensurate with the ExoMars localization
accuracy requirements, further development is required
to improve the accuracy and achieve robustness under a
variety of terrain conditions, and decreased
computational cost for greater practicality for a flight
mission.
The latest MDA VME approach is presented in this
paper. This extended localization system was tested in a
field campaign in the Mojave Desert using the robot
shown in . The new features being tested
which yield improvements from the previous work
include:
• Optimal use of different features from stereo-pair
images as visual landmarks. SIFT features are
invariant to image translation, scaling, rotation, and
partially invariant to illumination changes and affine
projection, making them suitable landmarks for use
in motion estimation. See [9] for details. However,
in testing to date of the VME SLAM system it has
been found that in some cases SIFT features alone
are insufficient for tracking. For example, on sandy
terrain with smooth-shaped rocks, the SIFT detector
often produces only small scale features, and these
are often unstable. Other features extracted from the
stereo-pair images are Maximally-Stable Extremal
Regions (MSERs) [10] and Harris-Laplace [11]
features.
• Investigation of the use of short-range active 3D
sensors (lighting condition independent) as the
visual front end to VME. A flash lidar-based, rigid
transform invariant 3D feature extraction technique
has been developed.
• Extension of the localization system to use
inclinometry to improve roll/pitch angle estimation.
Inclinometry maintains an observable estimate of the
gravity vector, instead of relying on roll and pitch
rate integration.
• Extension of the localization system to use a
long-range and wide FOV active 3D sensor in order
to extract long-range fixed landmarks for enforcing
VME observability, and thus improving the accuracy
of the approach.
• Development of an adaptive Extended Kalman Filter
(EKF) based SLAM
• Use of the improved VME estimated pose as
feedback for closed loop motion control. Prior work
in this field has focused on VME as an open-loop
observer, for providing accurate a posteriori
localization when a destination is reached. In our
novel approach, the low-rate and delayed VME
measurements and the high-rate IMU-odometry
measurements are fused to yield a high-rate and
smooth position and attitude measurement signal.
This technique is called “frequency lifting”. The
frequency lifted output is then used as the feedback
signal for closed loop path tracking
This new VME technology was integrated as part of
a complete short-range GN&C solution for planetary
rovers, including 3D terrain modeling and assessment,
motion planning and tracking, hazard detection and
avoidance algorithms, and high level supervisory control
based on CORTEX, a tool developed by the CSA for
design and real time execution of state machines [12].
The rest of this paper is organized as follow. Section
2 describes a typical field test scenario and provides the
field test terrain landscape through pictures and maps
reconstructed using sensor data. The challenges and the
issues observed during the field test campaign are also
presented. The experimental results are presented and
discussed in section 3, and the lessons learned are also
discussed. Section 4 concludes the paper.
2 Test Scenario and Field Test Terrain
Landscapes
This section describes a typical field test scenario of
AIR-GNC and provides the field test terrain landscapes
through pictures and maps reconstructed using sensor
data. The goal of AIR-GNC was to advance the
capabilities of autonomous terrain modeling and
assessment, and localization for future planetary
missions. In order to demonstrate these capabilities, it
was necessary to test them in a relevant planetary
environment, which includes rocks, crevices and cracks,
loose sand, and slope hazards.
The dry lake beds of the Mojave Desert in the
Southern United States have been used as Martian
554
analogues in the past by NASA’s Jet Propulsion
Laboratory [13]. The terrain offers a variety of soil
conditions, from dry fine grain sand, to hard packed clay,
to loose gravel. Slopes and rock fields can also be found
near the edges of the lake beds. The area is also relatively
free of vegetation.
The testbed with the newly developed software
modules can be fully exercised due to the presence of
slope and rock hazards, the likelihood of wheel slippage,
and the presence of representative natural image features.
2.1 The Challenges and Issues Observed
This subsection intends to address some of the
general issues observed which affected the field
campaign. Comments on the performance of the
localization system itself will be addressed in the results
section. General issues during the field campaign
included:
• Weather: January is the rainy season in the Mojave
Desert. Unfortunately, in this field campaign, it
rained more frequently and heavily than usual. As a
result, a few full days of testing were completely lost,
and about four days had to be cut short due to rain.
Rainfall could be heavy at times, flooding the dry
lake beds. With any more than a light sprinkle, the
playa becomes very slick and loose, quickly causing
a risk to the vehicles. Further, the ground needs time
to dry out after a heavy rainfall, meaning that more
than one day can be lost at a time.
• Glare: The glare from the sunlight on the moistened,
high-reflectivity sand frequently caused glare in the
VME stereo camera images. These specular
reflections had a negative impact on the
performance of the VME localization. For a more
complete discussion, see the results section.
• Rover Terrainability: The terrainability of the rover
(See ) used in the test campaign is very
limited. Having no suspension and limited ground
clearance means that very small rocks can be
hazardous. While the existing sensor suite certainly
has the resolution to detect hazards for a
planetary-representative chassis, with the Red Rover
there are still some undetected rocks which, while
not catastrophic, can cause the rover to stick and
skid, or undergo abrupt pitch and roll motions.
• Shadows in the images: No effort was made to
control the direction of travel with respect to the sun,
and as such there are many instances in which the
shadow of the rover appears in the stereo camera
images. This ensures that the motion filtering in the
feature matching algorithms is fully exercised.
2.2 Field Test Scenario
To better understand the sequence of operations
during the test campaign, this subsection describes a
typical long-range autonomous traverse scenario. This
scenario can be classified as an autonomous navigation
test in unknown terrain. The terrain is considered static,
although the system is equipped with dynamic obstacle
detection ability.
Typical Field test scenario, showing 3D
terrain data from the sensor, colour coded according
to traversability.
Typical operation (See ) starts with:
• A remotely located operator or scientist selects
goal locations that are outside of the rover’s
on-board sensor ranging envelope, and/or
obscured by an obstacle from a start location.
• The rover takes a high resolution medium range
(10 to 15 m) scan of its surrounding
environment using its integrated sensor suite.
• The rover constructs an internal representation
of the surrounding terrain, conducts the terrain
traversability assessment and plans a hazard free
path toward the goal. This path is safe inside the
scanned area, and aims directly at the goal
outside the sensed area.
• The operator station displays the terrain map
including traversability map overlays, and the
planned path.
• The rover tracks the planned path up to the
boundary of the medium range map and stops.
Telemetry and environment data are collected
during the motion and at the stop location.
• Mapping, terrain assessment, and path planning
555
are automatically repeated until the final
destination is reached. If the goal is not
reachable, the rover will move as close as
possible to the defined goal and request a new
goal or a new task.
All of the above steps are coordinated and monitored by
a high level supervisory control and data acquisition
module developed using the CSA’s CORTEX autonomy
framework.
2.3 Field Test Terrain Landscapes and
Reconstructed Maps
The field tests were conducted in and around Silver
Dry Lake, Silurian Dry Lake, the Mudflat, and Dumont
Little Dunes in the Mojave Desert. The test environs
offer many interesting terrains (in terms of rock, bush,
and slope hazards, visual features, soil characteristics) all
within a small geographical area.
Silver Dry Lake Test Locations
Three locations were selected for testing in the Silver
dry lake bed: the “playa”, the “boulder”, and the
“plateau” locations. The “playa” location is characterized
by relatively flat, open, and hazard-free terrain. The soil
is hard clay, which has cracked as it dries. This produces
a fractal-like pattern of fissures on the surface. The
fissures also make the surface bumpy, which can cause
the rover to shake as it moves (See ).
The “boulder” field is characterized by relatively flat
loose sand and playa, with large rock hazards randomly
scattered throughout. (See ).
The “plateau” area is characterized by loose, small
grain gravel with few rocks, but many medium-sized
bushes (See and ).
Typical view of the Playa location
Silurian Dry Lake Test Locations
Three locations were selected for testing in the
Silurian dry lake bed: the “playa”, the “bush island”, and
the “shore” locations. The open playa of Silurian lake is
virtually the same as the playa of Silver lake, with the
exception that the fissures in the clay are slightly larger.
The “bush island” location is in the middle of the
Silurian lake playa, and contains a wide variety of soil
conditions: hard cracked clay, soft cracked clay, gravel,
and sand. There are several medium-sized bushes and a
sparse scattering of rocks. There are also some mild
gulleys to provide interesting slopes. The “shore”
location is simply the shore of the Silurian dry lake bed,
where the playa turns into loose gravel with some bushes
(See ).
Typical view of the Boulder location
Mudflat Test Location
This area contains a lot of loose sand and gravel,
with many ravines and gulleys. There are only a few
bushes. Most of the hazards are due to slope rather than
rocks (See and ).
Typical view of the Plateau location
556
Reconstructed Digital Elevation Map of the
plateau and the traveled path (200 x 255m, 960 m
traveled distance)
Typical view of the Shore location
Typical view of the mudflat location
Reconstructed Mudflat Digital Elevation
Map and traveled path overlays (160m x 140m, 445 m
traveled distance)
Dumont Little Dunes Test Location
The terrain is loose sand dunes, with occasional
sparse bushes as shown in .
Typical view of the Dunes location
3 Experimental Results
This section describes the field trials of AIR-GNC
performed in the Mojave Desert of Southern California
in January, 2010.
3.1 Performance Metrics and Statistics
The Ground truth position was measured using the
NavCom 3020M real-time kinematic differential global
positioning system (RTK DGPS). The manufacturer’s
specification states an accuracy of 1 cm when operating
557
in RTK mode [14]. GPS latitude, longitude, and altitude
data were converted to the local frame where positive X
is East, positive Y is North, and positive Z is up (i.e.,
coincident with the local gravity direction).
Localization data was logged with respect to the
initial rover body frame at the start of the traverse. As
such, the ground truth must be expressed in the initial
rover frame before accuracy analyses can begin. The
position offset is easy to correct by simply subtracting
the initial GPS position from all subsequent
measurements. Pitch and roll are also compensated, since
the VME initializes itself using the inclinometry from the
IMU’s accelerometers, which give very clean
measurements since the rover is stationary. However, in
the absence of an absolute heading sensor on the rover,
the alignment of the ground truth becomes somewhat
more challenging.
In this project, with so many different options to the
VME algorithm, it was more desirable to align the
ground truth with the IMU-corrected odometry. Note that
no matter which signal is used for the alignment, the
accuracy results will be biased (even if very slightly) in
favour of that signal over the others. In the absence of
absolute attitude ground truth, this problem cannot be
avoided.
The distance between the end point of the estimate
and the end point of the ground truth, referred to as
end-point distance, is used to evaluate the localization
accuracy.
3.2 Result and Analysis
VME localization accuracy results are compared
based on the visual feature used for localization.
shows the normalized localization percentage error
by feature based on the data collected in all field test
locations. The width of each bin of the histogram is 2%
(of distance travelled), and for clarity, the axis is cut off
at 10% error. All runs with errors greater than 10% (more
likely outliers) are lumped into a single histogram bin,
which is presented to the right of the regular histogram.
As can be seen in the figure above, on average
MSER outperformed both Harris-Laplace and SIFT. This
suggests that in these terrains the image regions had
greater stability than the local (corner) features. On the
average, Enhanced IMU-corrected odometry had slightly
better accuracy than MSER. and
show the normalized localization result of respectively
the Enhanced IMU-corrected odometry and MSER
across test locations.
localization percentage error by feature
From and it can be seen that
MSER outperformed Enhanced IMU-corrected odometry
in the Playa area and the Dunes area, the latter certainly
due to the high amount of wheel slippage.
The Mudflat location was particularly difficult for every
localization method, including Enhanced IMU-corrected
odometry. In the latter case, the terrain was quite loose,
which caused considerable slippage. VME was primarily
affected by the high reflectivity of the terrain (for
example, the glare which was present in many runs).
Moreover, the stereo cameras used in this project for
VME have automatic brightness and contrast adjustment
capability. One of the lessons learned from this field trial
is that automatic exposure adjustment may actually
hinder the performance of the stereo camera vision
system. In many runs there is a significant and
sometimes intense glare from the sunlight, which washes
out large portions of the image and increases the contrast.
Example runs in which the glare is very prominent
include those on the Playa, and in the Mudflat location
before the sun goes behind the mountains (See
). Further, this glare is specular, so it changes with the
relative angle to the sun. As such, the glare patches
themselves do not make very reliable features as they
change with viewpoint. This effect was not noticed
during the previous field campaign in 2008, which took
place during the dry season [8].
558
Enhanced IMU-corrected odometry
localization percentage across all test locations
MSER localization percentage across all
test locations.
3.3 Summary of the results
summarizes the localization accuracy of
stereo camera feature detectors shown in .
Feature Min % Error Mean % Error Std. Dev
Harris-Lap. 0.8 7.3 5.2
MSER 0.4 3.9 2.8
SIFT 1.2 8.8 11.9
IMU-Odo 0.4 3.5 2.2
MSER had better accuracy and lower standard
deviation than both Harris-Laplace and SIFT. This
suggests that in these types of terrains, region-based
features are more useful than the local (corner) features.
This is particularly noticeable in the Playa areas. MDA
Enhanced IMU-corrected odometry did very well in most
locations; slightly better than MSER. MSER
outperformed Enhanced IMU-corrected odometry in the
Playa and Dunes areas. Harris-Laplace performed nearly
as well as Enhanced IMU-corrected odometry in the
Shore area (the plot is not shown). The Mudflat location
was particularly difficult for all methods of localization.
As mentioned above, Enhanced IMU-corrected odometry
was subject to slippage, and camera images were subject
to glare.
Based on above results, an elegant localization
solution will rely on Enhanced IMU-corrected odometry
and continuous monitoring for wheel slippage. If the
slippage starts, the localization will switch from
Enhanced IMU-corrected odometry to VME and /or
map-based localization and stay in this mode until the
slippage ends. Another solution is to robustly extract and
match long-range landmarks online to anable the
observability of EKF-based SLAM. Offline results (See
) using data collected during the field trials have
shown very promising results.
Location Distance IMU-Odo MSER LRO
Mudflat, run 1 85 m 2.0% 3% 0.6%
Mudflat, run 2 21 m 3.0% 9.4% 0.9%
Plateau 34 m 1.5% 5.3% 1.3%
Mudflat terrain - High contrast due to
glare off of the moist mudflat terrain
559
4 Conclusions and Future work
This paper has presented the field test results of the
planetary VME for AIR-GNC project conducted by
MDA in collaboration with and funded by the CSA.
The operational scenario of AIR-GNC, which
includes terrain scanning, modeling and traversability
assessment, and path planning and tracking, has been
tested to a high level of maturity. CORTEX-based
supervisory control was effective for monitoring and
executing traverses, which resulted in a noticeable
increase in the autonomy of the testbed. An added
advantage is that experiments could be set up and run in
a fraction of the time compared to the 2008 field
campaign. VME-based feedback was successfully used
to close the path tracking loop.
Enhanced IMU-corrected odometry performed very
well; on average it performed slightly better than MSER
based VME. MSER outperformed Harris-Laplace, and
SIFT as a feature extraction front end to VME.
Future work includes wheel slippage monitoring and
estimation using visual evidence, as discussed in section
3.3. This improved odometry could potentially be
combined with state estimation using a dynamic model
of the rover, and would be a wide-ranging benefit. For
the visual front end, a fast and robust long range
3D-feature extraction method will be critical in order to
take full advantage of observability.
5 Acknowledgment
This work was funded by the Canadian Space
Agency. The authors would like to thank Régent
L’Archeveque and David Gingras from the CSA for their
valuable contribution in path planning and CORTEX
based supervisory control. The authors also thank
Manickam Umasuthan from MDA for his contributions
to the 3D feature extraction implementation, and Tung
Nguyen and Kelvin Tao, students at the University of
Toronto, for integration of sensors and initial testing of
the system.
References
[1] R. Volpe, Rover functional autonomy development
for the Mars mobile science laboratory. In
Proceedings of the IEEE Aerospace Conference,
volume 2, pages 643–652, Big Sky, MT, USA, 2006.
[2] M. Maimone, Y. Chang, and L. Matthies, “Two years
of visual odometry on the Mars exploration rovers”.
Journal of Field Robotics, 3(24):169–286, 2007.
[3] S. Lacroix, et al. “Autonomous Rover Navigation on
Unknown Terrains. Demonstrations in the Space
Museum Cité de l'Espace at Toulouse”. In 7th Int.
Symposium on Experimental Robotics. Honolulu, HI
(USA), 2000.
[4] S. Lacroix S. Lacroix, et al.. “Autonomous Rover
Navigation on Unknown Terrains: Functions and
Integration”, In International Journal of Robotics
Research, 21(10-11), pages 917-942, 2002.
[5] P.I. Corke, et al. “Omnidirectional visual odometry
for a planetary rover”. In Proceedings of Intelligent
Robot and System, 2004
[6] J.J. Biesiadecki,et al., “Mars exploration rover
surface operations: Driving opportunity at meridiani
planum”. In IEEE Conference on Systems, Man and
Cybernetics, The Big Island, Hawaii, USA, 2005.
[7] M.W. Maimone, Y. Cheng, L. Matthies, “Two years
of visual odometry on the mars exploration rovers:
Field reports”, J. Field Robot., 24(3):169-186, 2007.
[8] J.N. Bakambu, C. Langley, and R. Mukherji, “Visual
Motion Estimation: Localization Performance
Evaluation Tool for Planetary Rovers”, In
International Symposium on Artificial Intelligence,
Robotics and Automation in Space (iSAIRAS), 2008.
[9] S. Se, T. Barfoot, P. Jasiobedzki, “Visual Motion
Estimation and Terrain Modeling for Planetary
Rovers”, iSAIRAS 2005, Munich, Germany.
[10] F. Kristensen, W.J. MacLean, “Real-time Extraction
of Maximally Stable Extremal Regions on an FPGA”,
IEEE Int. Symp. Circuits and Systems, 2007.
[11] K. Mikolajczyk, et al., “A Comparison of Affine
Region Detectors”, IJCV, 65(1-2): 43-72, 2005.
[12] E. Dupuis, P. Allard and R. L’Archevêque,
“Autonomous Robotics Toolbox”, iSAIRAS, Munich,
5-8 September 2005.
[13] T. Huntsberger, et al, “Rover autonomy for long
range navigation and science data acquisition on
planetary surfaces”, IEEE ICRA 2002.
[14] NavCom Technology Inc., “RT-3020 GPS Products
User Guide”, 2008.